[FFmpeg-devel] [PATCH v2 1/3] hwcontext_vulkan: do not use optical flow queueus by default

[FFmpeg-devel] [PATCH v2 1/3] hwcontext_vulkan: do not use optical flow queueus by default

[Lynne]

We don't use them, and on NVIDIA, each queue takes around 30ms
to allocate, and the driver has a global limit of ONLY 112 queues.
---
 libavutil/hwcontext_vulkan.c | 8 --------
 1 file changed, 8 deletions(-)

diff --git a/libavutil/hwcontext_vulkan.c b/libavutil/hwcontext_vulkan.c
index ce485a85a2..dc9d49ccb4 100644
--- a/libavutil/hwcontext_vulkan.c
+++ b/libavutil/hwcontext_vulkan.c
@@ -95,8 +95,6 @@ typedef struct VulkanDeviceFeatures {
     VkPhysicalDeviceDescriptorBufferFeaturesEXT descriptor_buffer;
     VkPhysicalDeviceShaderAtomicFloatFeaturesEXT atomic_float;
 
-    VkPhysicalDeviceOpticalFlowFeaturesNV optical_flow;
-
 #ifdef VK_KHR_shader_relaxed_extended_instruction
     VkPhysicalDeviceShaderRelaxedExtendedInstructionFeaturesKHR relaxed_extended_instruction;
 #endif
@@ -234,9 +232,6 @@ static void device_features_init(AVHWDeviceContext *ctx, VulkanDeviceFeatures *f
     FF_VK_STRUCT_EXT(s, &feats->device, &feats->relaxed_extended_instruction, FF_VK_EXT_RELAXED_EXTENDED_INSTR,
                      VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_RELAXED_EXTENDED_INSTRUCTION_FEATURES_KHR);
 #endif
-
-    FF_VK_STRUCT_EXT(s, &feats->device, &feats->optical_flow, FF_VK_EXT_OPTICAL_FLOW,
-                     VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_OPTICAL_FLOW_FEATURES_NV);
 }
 
 /* Copy all needed device features */
@@ -311,7 +306,6 @@ static void device_features_copy_needed(VulkanDeviceFeatures *dst, VulkanDeviceF
     COPY_VAL(expect_assume.shaderExpectAssume);
 #endif
 
-    COPY_VAL(optical_flow.opticalFlow);
 #undef COPY_VAL
 }
 
@@ -607,7 +601,6 @@ static const VulkanOptExtension optional_device_exts[] = {
     { VK_EXT_PHYSICAL_DEVICE_DRM_EXTENSION_NAME,              FF_VK_EXT_DEVICE_DRM             },
     { VK_EXT_SHADER_ATOMIC_FLOAT_EXTENSION_NAME,              FF_VK_EXT_ATOMIC_FLOAT           },
     { VK_KHR_COOPERATIVE_MATRIX_EXTENSION_NAME,               FF_VK_EXT_COOP_MATRIX            },
-    { VK_NV_OPTICAL_FLOW_EXTENSION_NAME,                      FF_VK_EXT_OPTICAL_FLOW           },
     { VK_EXT_SHADER_OBJECT_EXTENSION_NAME,                    FF_VK_EXT_SHADER_OBJECT          },
     { VK_KHR_SHADER_SUBGROUP_ROTATE_EXTENSION_NAME,           FF_VK_EXT_SUBGROUP_ROTATE        },
 #ifdef VK_KHR_shader_expect_assume
@@ -1511,7 +1504,6 @@ static int setup_queue_families(AVHWDeviceContext *ctx, VkDeviceCreateInfo *cd)
     PICK_QF(VK_QUEUE_GRAPHICS_BIT, VK_VIDEO_CODEC_OPERATION_NONE_KHR);
     PICK_QF(VK_QUEUE_COMPUTE_BIT, VK_VIDEO_CODEC_OPERATION_NONE_KHR);
     PICK_QF(VK_QUEUE_TRANSFER_BIT, VK_VIDEO_CODEC_OPERATION_NONE_KHR);
-    PICK_QF(VK_QUEUE_OPTICAL_FLOW_BIT_NV, VK_VIDEO_CODEC_OPERATION_NONE_KHR);
 
     PICK_QF(VK_QUEUE_VIDEO_ENCODE_BIT_KHR, VK_VIDEO_CODEC_OPERATION_ENCODE_H264_BIT_KHR);
     PICK_QF(VK_QUEUE_VIDEO_DECODE_BIT_KHR, VK_VIDEO_CODEC_OPERATION_DECODE_H264_BIT_KHR);
-- 
2.49.0.395.g12beb8f557c
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel

To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".

[FFmpeg-devel] [PATCH v2 2/3] hwcontext_vulkan: minimize queue allocation on NVIDIA

[Lynne]

On NVIDIA, there's a global maximum limit of approximately 112 queues,
which means it takes ONLY 7 total programs using the maximum amount of
queues to cause the driver to error out/*segfault* during initialization.

Also, each queue takes about 30ms to allocate, which quickly adds up.

This reduces the queues allocate to the minimum that we would be happy
with. Its not worth limiting decode/encode queues as they're generally
not a lot, and do help.
---
 libavutil/hwcontext_vulkan.c | 24 +++++++++++++++++++++++-
 1 file changed, 23 insertions(+), 1 deletion(-)

diff --git a/libavutil/hwcontext_vulkan.c b/libavutil/hwcontext_vulkan.c
index dc9d49ccb4..243d68dc44 100644
--- a/libavutil/hwcontext_vulkan.c
+++ b/libavutil/hwcontext_vulkan.c
@@ -1418,6 +1418,13 @@ static int setup_queue_families(AVHWDeviceContext *ctx, VkDeviceCreateInfo *cd)
     VulkanDevicePriv *p = ctx->hwctx;
     AVVulkanDeviceContext *hwctx = &p->p;
     FFVulkanFunctions *vk = &p->vkctx.vkfn;
+    VkPhysicalDeviceDriverProperties dprops = {
+        .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES,
+    };
+    VkPhysicalDeviceProperties2 props2 = {
+        .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROPERTIES_2,
+        .pNext = &dprops,
+    };
 
     VkQueueFamilyProperties2 *qf = NULL;
     VkQueueFamilyVideoPropertiesKHR *qf_vid = NULL;
@@ -1471,7 +1478,14 @@ static int setup_queue_families(AVHWDeviceContext *ctx, VkDeviceCreateInfo *cd)
 
     hwctx->nb_qf = 0;
 
-    /* Pick each queue family to use */
+    /* NVIDIA's proprietary drivers have stupid limits, where each queue
+     * you allocate takes tens of milliseconds, and the more queues you
+     * allocate, the less you'll have left before initializing a device
+     * simply fails (112 seems to be the max). GLOBALLY.
+     * Detect this, and minimize using queues as much as possible. */
+    vk->GetPhysicalDeviceProperties2(hwctx->phys_dev, &props2);
+
+    /* Pick each queue family to use. */
 #define PICK_QF(type, vid_op)                                            \
     do {                                                                 \
         uint32_t i;                                                      \
@@ -1495,6 +1509,14 @@ static int setup_queue_families(AVHWDeviceContext *ctx, VkDeviceCreateInfo *cd)
         if (i == hwctx->nb_qf) {                                         \
             hwctx->qf[i].idx = idx;                                      \
             hwctx->qf[i].num = qf[idx].queueFamilyProperties.queueCount; \
+            if (dprops.driverID == VK_DRIVER_ID_NVIDIA_PROPRIETARY) {    \
+                if (type == VK_QUEUE_GRAPHICS_BIT)                       \
+                    hwctx->qf[i].num = FFMIN(hwctx->qf[i].num, 1);       \
+                if (type == VK_QUEUE_COMPUTE_BIT)                        \
+                    hwctx->qf[i].num = FFMIN(hwctx->qf[i].num, 4);       \
+                if (type == VK_QUEUE_TRANSFER_BIT)                       \
+                    hwctx->qf[i].num = FFMIN(hwctx->qf[i].num, 2);       \
+            }                                                            \
             hwctx->qf[i].flags = type;                                   \
             hwctx->qf[i].video_caps = vid_op;                            \
             hwctx->nb_qf++;                                              \
-- 
2.49.0.395.g12beb8f557c
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel

To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".

[FFmpeg-devel] [PATCH v2 3/3] hwcontext_vulkan: add a setting to limit queues

[Lynne]

If its a problem, you'll likely want to set it to 1 than more fine-grained
control, which you can already do via the API.
---
 libavutil/hwcontext_vulkan.c | 21 +++++++++++++++++----
 1 file changed, 17 insertions(+), 4 deletions(-)

diff --git a/libavutil/hwcontext_vulkan.c b/libavutil/hwcontext_vulkan.c
index 243d68dc44..2d3a95d263 100644
--- a/libavutil/hwcontext_vulkan.c
+++ b/libavutil/hwcontext_vulkan.c
@@ -143,6 +143,9 @@ typedef struct VulkanDevicePriv {
     /* Disable multiplane images */
     int disable_multiplane;
 
+    /* Maximum queues */
+    int limit_queues;
+
     /* Nvidia */
     int dev_is_nvidia;
 } VulkanDevicePriv;
@@ -1509,13 +1512,18 @@ static int setup_queue_families(AVHWDeviceContext *ctx, VkDeviceCreateInfo *cd)
         if (i == hwctx->nb_qf) {                                         \
             hwctx->qf[i].idx = idx;                                      \
             hwctx->qf[i].num = qf[idx].queueFamilyProperties.queueCount; \
-            if (dprops.driverID == VK_DRIVER_ID_NVIDIA_PROPRIETARY) {    \
+            if (p->limit_queues ||                                       \
+                dprops.driverID == VK_DRIVER_ID_NVIDIA_PROPRIETARY) {    \
+                int max = p->limit_queues;                               \
                 if (type == VK_QUEUE_GRAPHICS_BIT)                       \
-                    hwctx->qf[i].num = FFMIN(hwctx->qf[i].num, 1);       \
+                    hwctx->qf[i].num = FFMIN(hwctx->qf[i].num,           \
+                                             max ? max : 1);             \
                 if (type == VK_QUEUE_COMPUTE_BIT)                        \
-                    hwctx->qf[i].num = FFMIN(hwctx->qf[i].num, 4);       \
+                    hwctx->qf[i].num = FFMIN(hwctx->qf[i].num,           \
+                                             max ? max : 4);             \
                 if (type == VK_QUEUE_TRANSFER_BIT)                       \
-                    hwctx->qf[i].num = FFMIN(hwctx->qf[i].num, 2);       \
+                    hwctx->qf[i].num = FFMIN(hwctx->qf[i].num,           \
+                                             max ? max : 2);             \
             }                                                            \
             hwctx->qf[i].flags = type;                                   \
             hwctx->qf[i].video_caps = vid_op;                            \
@@ -1722,6 +1730,11 @@ static int vulkan_device_create_internal(AVHWDeviceContext *ctx,
     if (opt_d)
         p->use_linear_images = strtol(opt_d->value, NULL, 10);
 
+    /* Limit queues to a given number if needed */
+    opt_d = av_dict_get(opts, "limit_queues", NULL, 0);
+    if (opt_d)
+        p->limit_queues = strtol(opt_d->value, NULL, 10);
+
     /* The disable_multiplane argument takes precedent over the option */
     p->disable_multiplane = disable_multiplane;
     if (!p->disable_multiplane) {
-- 
2.49.0.395.g12beb8f557c
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel

To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".